Boost C++ Libraries: Ticket #12411: Boost coroutines 1.61 do not preserve fpu registers with O2 optimizations in Visual Studio

olli — Thu, 25 Aug 2016 05:56:10 GMT

could you provide code that demonstrates the problem, please

attachment set

martin_haas@… — Fri, 26 Aug 2016 20:09:15 GMT

attachment → BoostCoroutineTest.cpp

Source code to show coroutine issue

martin_haas@… — Fri, 26 Aug 2016 20:16:16 GMT

Please see the attached file. When compiled with Visual Studio 2015, the Debug versions work correctly. The release version works correctly if optimization is disabled. If optimization is specified as O2 or Ox, it appears the fpu registers are being cleared. In the working versions, the final job completes at time 54. In the nonworking versions, the final job appears to take zero time, so it completes at time 44.

if the "hold" function is changed to return an int instead of a double, the optimized programs work correctly.

The program works correctly with boost 1.60, but not 1.61.

Thanks for all your hard work. The performance of the coroutines are MUCH better than my current solution. Hope this is an easy fix.

Thanks

olli — Sun, 28 Aug 2016 08:35:23 GMT

unfortunately, boost.coroutine is deprecated and will not be fixed or developed further. You coud use boost.context, boost.coroutine2 (no symmetric coroutines) or boot.fiber instead.

For the problem you mentioned - could you provide a simple example using boost.context that demonstrates problems with the FPU env, please?

component changed

olli — Sun, 27 Nov 2016 08:45:15 GMT

component coroutine → context

severity changed

olli — Sun, 27 Nov 2016 08:45:57 GMT

severity Showstopper → Problem

status changed; resolution set

olli — Mon, 05 Dec 2016 07:40:28 GMT

status new → closed
resolution → fixed

martin_haas@… — Mon, 05 Dec 2016 16:50:33 GMT

Replying to olli: Thanks Oliver. I verified that the change worked for me. Will this change in Boost Context be rolled into the Boost 1.63 release?

Thanks

anonymous — Mon, 05 Dec 2016 20:02:01 GMT

Replying to martin_haas@…:

Will this change in Boost Context be rolled into the Boost 1.63 release?

yes